Deep learning driven inverse design of multi-material structures with tailored anisotropic mechanical responses

Multi-material structures are complex architectures formed by the selective distribution of diverse materials at the microstructural level, enabling the anisotropic designs with customized mechanical responses. Although these structures offer significant advantages via their immense design freedom, the vast design space and the non-linear coupling of multi-phase mechanical properties pose formidable challenges for inverse design of such structures. This study develops a deep learning-driven framework for the rapid design of voxelized multi-material structures. By integrating a high-fidelity forward surrogate model with an inverse neural network, our method can generate spatial material distributions that satisfy targeted stress-strain curves within seconds. Validated through multi-material 3D printing and quasi-static compression tests, the framework achieves an accuracy of over 95%. Notably, we introduce a joint loss function coupled with a Hard Constraint Check (HCC) strategy, allowing the model to selectively bias designs toward specific soft-to-hard material ratios without compromising mechanical performance. Furthermore, a weighted multi-objective optimization scheme is implemented to incorporate priorities when managing anisotropic responses. Experimental results demonstrate the immense potential of deep learning in the spatial distribution design of multi-material structures, paving the way for advanced material applications in fields such as medicine, robotics, aerospace, civil engineering, and vehicle engineering.